Read enough writing and you develop a sense for it. A paragraph can be grammatically flawless and still feel hollow. Another can be full of small errors and feel completely alive.
The difference is not quality in the usual sense. It is a set of structural habits that separate text generated one token at a time from text produced by someone who had something to say.
Human writing is uneven. A person writing naturally produces a three-word sentence, then a thirty-word sentence, then a fragment. The rhythm changes because the thought changes.
AI text tends to move at a steady pace. Paragraphs run similar lengths, sentences cluster around the same complexity, and transitions appear at regular intervals. It reads smoothly, and smoothness over hundreds of words is itself a signal.
This is what researchers call burstiness: the amount of variation in sentence length and complexity. Human writing is high in burstiness. Model output is low.
People assume AI betrays itself through vocabulary. The more reliable tell is structure.
Ask a person to write about three benefits of something and they will usually organize the three however the ideas came to them. Ask a model and you often get three paragraphs with the same internal shape: a topic sentence, an elaboration, an example, a summary line.
Three parallel paragraphs is not proof of anything. Three parallel paragraphs with identical internal geometry is a pattern worth noticing.
Human writers reach for details from their own life because those details are free and close at hand. A model reaches for the most probable detail, which is usually the most generic one.
Compare "our support team cut response times by routing tickets by product line" with "our support team improved response times through strategic optimizations." The first could only be written by someone who was there.
Specificity is the hardest thing to fake and the easiest thing to add. If a draft feels hollow, the fix is almost always a number, a name, or a story.
Human writers hedge unevenly. They are confident about what they know and openly unsure about what they do not. AI text tends to hedge uniformly, sprinkling qualifiers everywhere, because it has no real confidence to express.
The reverse pattern also shows up: a model states an invented fact with total assurance while carefully qualifying something obvious.
Human writers use transitions sparingly and often informally. Model output loves the full repertoire: furthermore, moreover, additionally, in conclusion, it is important to note.
One or two of these in a long document is normal. A transition at the start of every paragraph is a stylistic fingerprint.
People make mistakes of attention: a repeated word, a dropped comma, a sentence that trails off. Models make mistakes of fact, and they make them fluently.
That distinction matters. A typo is human evidence. A confident fabricated statistic is not a typo, it is a different failure entirely.
Human writing usually carries evidence of having been thought about. There is an aside that goes nowhere. A metaphor is started and abandoned. An example is overexplained because the writer found it interesting.
Model output has no abandoned threads, because nothing was ever genuinely pursued. Every sentence serves the assigned task.
You do not need a detector to notice the pattern, though a detector helps. Try these:
It is worth being fair to the tool. Models are excellent at structure, at producing a usable first draft, and at getting past a blank page. They are bad at specificity and at having a view.
The most effective way to use one is to write the parts only you can write, then use the model for the scaffolding around them. A human outline plus a model draft plus a human rewrite is a normal 2026 workflow, and it produces text that reads as human because the thinking in it is human.
A detector is not reading for soul. It is estimating how probable each successive word is given what came before. Human writing is, statistically, less predictable. That is the whole mechanism, and it explains every pattern above.
It also explains the limit. A careful human can write predictably. A model can be prompted toward variety. Neither changes what the text fundamentally is, and both are reasons to treat any single score as one signal among several.